Measuring Data Bias, Part 2: Comparative Analysis
This part of the assignment goes deeper into understanding how open-source and closed-source generative AI models handle data bias. You will compare these models to see how each responds to similar prompts and whether certain content is blocked or restricted.
Instructions:
Select an open-source model and a closed-source image or text generation model.
Compare the two models and experiment to see whether you can increase the bias of generations. Test the closed-source model to see what it generates in similar scenarios and what types of content it blocks from generating, if any. Based on your investigation, what kind of mitigations do you think might be in place for the closed-source systems, and why?
Deliverables:
For part 2, please submit:
The results on your evaluation set from Part 1 from both the open-source and the closed-source models.
A minimum of 2-3 new contexts that demonstrate a difference in bias mitigation between the open-source and closed-source models (or, if no such difference materializes, include the new contexts and briefly describe the responses for both).
A 500-word writeup of your findings when comparing the two systems on your bias evaluation set, including:
Any evidence for or hypotheses about specific mitigations that either system might have
Interesting outputs or surprising findings, if any, surfaced when probing for bias
Your overall assessment of bias and mitigation as based on the combined results from Parts 1 and 2.